Message ID | 20200108001859.25254-1-sean.j.christopherson@intel.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | KVM: x86/mmu: Fix a benign Bitwise vs. Logical OR mixup | expand |
Sean Christopherson <sean.j.christopherson@intel.com> writes: > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > checks, which are obviously intended to be logical statements. Switching > to a Logical OR is functionally a nop, but allows the compiler to better > optimize the checks. > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > --- > arch/x86/kvm/mmu/mmu.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > index 7269130ea5e2..72e845709027 100644 > --- a/arch/x86/kvm/mmu/mmu.c > +++ b/arch/x86/kvm/mmu/mmu.c > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > { > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); Redundant parentheses detected! > } Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
On Wed, Jan 8, 2020 at 2:13 AM Vitaly Kuznetsov <vkuznets@redhat.com> wrote: > > Sean Christopherson <sean.j.christopherson@intel.com> writes: > > > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > > checks, which are obviously intended to be logical statements. Switching > > to a Logical OR is functionally a nop, but allows the compiler to better > > optimize the checks. > > > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > > --- > > arch/x86/kvm/mmu/mmu.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > > index 7269130ea5e2..72e845709027 100644 > > --- a/arch/x86/kvm/mmu/mmu.c > > +++ b/arch/x86/kvm/mmu/mmu.c > > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > > { > > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); > > Redundant parentheses detected! I think you mean superfluous rather than redundant. > > } > > Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com> Reviewed-by: Jim Mattson <jmattson@google.com>
From: Sean Christopherson > Sent: 08 January 2020 00:19 > > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > checks, which are obviously intended to be logical statements. Switching > to a Logical OR is functionally a nop, but allows the compiler to better > optimize the checks. > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > --- > arch/x86/kvm/mmu/mmu.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > index 7269130ea5e2..72e845709027 100644 > --- a/arch/x86/kvm/mmu/mmu.c > +++ b/arch/x86/kvm/mmu/mmu.c > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > { > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); Are you sure this isn't deliberate? The best code almost certainly comes from also removing the '!= 0'. You also don't want to convert the expression result to zero. So: return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | (rsvd_check->bad_mt_xwr & (1ull << low6)); The code then doesn't have any branches to get mispredicted. David - Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK Registration No: 1397386 (Wales)
On Thu, Jan 09, 2020 at 02:13:48PM +0000, David Laight wrote: > From: Sean Christopherson > > Sent: 08 January 2020 00:19 > > > > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > > checks, which are obviously intended to be logical statements. Switching > > to a Logical OR is functionally a nop, but allows the compiler to better > > optimize the checks. > > > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > > --- > > arch/x86/kvm/mmu/mmu.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > > index 7269130ea5e2..72e845709027 100644 > > --- a/arch/x86/kvm/mmu/mmu.c > > +++ b/arch/x86/kvm/mmu/mmu.c > > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > > { > > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); > > Are you sure this isn't deliberate? > The best code almost certainly comes from also removing the '!= 0'. > You also don't want to convert the expression result to zero. The function is static inline bool, so it's almost certainly a mistake originally. The != 0 is superfluous, but this will get inlined anyway. > > So: > return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | (rsvd_check->bad_mt_xwr & (1ull << low6)); > The code then doesn't have any branches to get mispredicted. > > David > > - > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > Registration No: 1397386 (Wales) >
On Thu, Jan 09, 2020 at 10:26:30AM -0500, Arvind Sankar wrote: > On Thu, Jan 09, 2020 at 02:13:48PM +0000, David Laight wrote: > > From: Sean Christopherson > > > Sent: 08 January 2020 00:19 > > > > > > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > > > checks, which are obviously intended to be logical statements. Switching > > > to a Logical OR is functionally a nop, but allows the compiler to better > > > optimize the checks. > > > > > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > > > --- > > > arch/x86/kvm/mmu/mmu.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > > > index 7269130ea5e2..72e845709027 100644 > > > --- a/arch/x86/kvm/mmu/mmu.c > > > +++ b/arch/x86/kvm/mmu/mmu.c > > > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > > > { > > > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > > > > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > > > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > > > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); > > > > Are you sure this isn't deliberate? > > The best code almost certainly comes from also removing the '!= 0'. The '!= 0' is truly superfluous, removing it doesn't affect code generation. > > You also don't want to convert the expression result to zero. > > The function is static inline bool, so it's almost certainly a mistake > originally. The != 0 is superfluous, but this will get inlined anyway. Ya, the bitwise-OR was added in commit 25d92081ae2f ("nEPT: Add nEPT violation/misconfigration support"), and AFAICT it's unintentional. That being said, I was a bit hasty in stating that a logical-OR allows for better optimization, sort of. For FNAME(prefetch_invalid_gpte) and FNAME(walk_addr_generic), which branch on the result of is_rsvd_bits_set(), the logical-OR is marginally better. FNAME(prefetch_invalid_gpte) is what I initially looked at when saying "yep, that's better!". But for walk_shadow_page_get_mmio_spte(), because it aggregates the result in a loop, the bitwise-OR is better in that it eliminates a Jcc. And all that being said, there are two vastly superior optimizations that can be made: - Reorder the checks in FNAME(prefetch_invalid_gpte) to perform the !PRESENT and !ACCESSED checks before checking the reserved bits, as they are both more likely to fail and do not require additional memory accesses. - Rewrite __is_rsvd_bits_set() to make it templated. The reserved MT check is EPT only, i.e. bad_mt_xwr is always 0 for legacy 32/64-bit paging. So, I'll scrap this patch and send a mini series to effect the above optimizations. > > > > So: > > return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | (rsvd_check->bad_mt_xwr & (1ull << low6)); > > The code then doesn't have any branches to get mispredicted. > > > > David > > > > - > > Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK > > Registration No: 1397386 (Wales) > >
On Thu, Jan 09, 2020 at 08:36:24AM -0800, Sean Christopherson wrote: > On Thu, Jan 09, 2020 at 10:26:30AM -0500, Arvind Sankar wrote: > > On Thu, Jan 09, 2020 at 02:13:48PM +0000, David Laight wrote: > > > From: Sean Christopherson > > > > Sent: 08 January 2020 00:19 > > > > > > > > Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit > > > > checks, which are obviously intended to be logical statements. Switching > > > > to a Logical OR is functionally a nop, but allows the compiler to better > > > > optimize the checks. > > > > > > > > Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> > > > > --- > > > > arch/x86/kvm/mmu/mmu.c | 2 +- > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c > > > > index 7269130ea5e2..72e845709027 100644 > > > > --- a/arch/x86/kvm/mmu/mmu.c > > > > +++ b/arch/x86/kvm/mmu/mmu.c > > > > @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) > > > > { > > > > int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; > > > > > > > > - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | > > > > + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || > > > > ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); > > > > > > Are you sure this isn't deliberate? > > > The best code almost certainly comes from also removing the '!= 0'. > > The '!= 0' is truly superfluous, removing it doesn't affect code > generation. Actually, it's not completely superfluous. Functionally the code is identical, but ordered slightly differently for whatever reason.
On 09/01/20 17:36, Sean Christopherson wrote: >>> You also don't want to convert the expression result to zero. >> The function is static inline bool, so it's almost certainly a mistake >> originally. The != 0 is superfluous, but this will get inlined anyway. > Ya, the bitwise-OR was added in commit 25d92081ae2f ("nEPT: Add nEPT > violation/misconfigration support"), and AFAICT it's unintentional. It may not be intentional in this case, but it's certainly the kind of code that I would have fun writing. :) Paolo
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c index 7269130ea5e2..72e845709027 100644 --- a/arch/x86/kvm/mmu/mmu.c +++ b/arch/x86/kvm/mmu/mmu.c @@ -3970,7 +3970,7 @@ __is_rsvd_bits_set(struct rsvd_bits_validate *rsvd_check, u64 pte, int level) { int bit7 = (pte >> 7) & 1, low6 = pte & 0x3f; - return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) | + return (pte & rsvd_check->rsvd_bits_mask[bit7][level-1]) || ((rsvd_check->bad_mt_xwr & (1ull << low6)) != 0); }
Use a Logical OR in __is_rsvd_bits_set() to combine the two reserved bit checks, which are obviously intended to be logical statements. Switching to a Logical OR is functionally a nop, but allows the compiler to better optimize the checks. Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com> --- arch/x86/kvm/mmu/mmu.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)